Gemini 4 Neon vs Carbon: when does Neon actually cost less?
Google’s November 18 launch prices Gemini 4 Neon below Carbon’s standard rate. Carbon’s temporary discount flips the comparison until December 1. Here is how to decide.
The cheaper model changes on December 1
Google introduced Gemini 4 Neon on November 18, 2026, and began making it available through the paid Gemini Developer API. The published rate is $3 per million input tokens and $15 per million output tokens. Carbon costs $2 and $10 at its temporary rates through November 30. Google lists December 1 as the date those Carbon rates rise to $4 and $20. For the same token usage, Neon therefore has a 50% premium over discounted Carbon and costs 25% less than Carbon at its scheduled standard rates.
That reversal matters more to a US team budgeting an API workflow than an isolated benchmark rank. Try Neon now on workloads where faster completion or better output could outweigh the promotional price gap. Set a reminder to revisit the routing decision once the posted Carbon rates change. Check the billing terms for your own account before moving production volume.
A fixed-token example isolates the price change
Suppose one task uses 100,000 uncached input tokens and 20,000 billable output tokens. Multiply 0.10 by the input rate and 0.02 by the output rate. The resulting token bill is $0.60 on Neon, $0.40 on discounted Carbon, and $0.80 on Carbon at its listed December rates. This is an illustration using identical token counts. It excludes cache effects, tool charges, retries, taxes and negotiated pricing.
Real workloads can use different amounts of output and thinking tokens. In particular, an agent that retries a failing step may use more tokens on the cheaper model. For an internal cost review, measure what you paid for the whole job, including failed attempts, then divide by correctly completed jobs. Record the thinking setting and any cache usage so that a retest in December can explain why the bill changed.
What the November benchmarks measured
Artificial Analysis reported Intelligence Index scores of 61 for Neon at xhigh and 60 for Carbon at high. It reported an estimated $6.44 versus $7.76 per benchmark task using the non-discounted Carbon price, with both models configured at those stated effort levels. The same report cites roughly 21 versus 36 minutes per task in that suite. These figures describe the evaluator’s setup. They do not tell you how long your requests will take or establish a one-point quality gain for your own workload.
Google’s launch post separately describes Neon as suited to complex coding and analytical work, while positioning Carbon for its hardest long-running projects. Treat that framing as a starting point for workload selection. For a representative trial, give both models the same inputs, access to the same tools and the same time budget. Use one acceptance rubric and include at least a few known hard cases where the current model still fails.
Who should try Neon first
A team that pays repeatedly for long document analysis or multi-file coding work is a good candidate for a trial, especially where latency delays a human review. Keep an existing Carbon route for work whose quality threshold Neon has yet to meet. Start by routing a limited share of comparable jobs to Neon and record total billable input and output tokens, time to accepted result, completion rate and rework minutes. A model switch is justified when it improves the team's real cost or delivery time at an acceptable error rate. BuzzWhy did not run this benchmark. The price example uses public rates and the reported performance figures come from Artificial Analysis.